Qualcomm AI Engine Direct - Quantizer refine for qat

Differential Revision: D65738212

Pull Request resolved: #6747

Author: chunit-quic

backends/qualcomm/quantizer/utils.py renamed to backends/qualcomm/quantizer/annotators.py

@@ -6,14 +6,12 @@
 from typing import Sequence
 
 import torch
+from executorch.backends.qualcomm.quantizer.annotators import QUANT_ANNOTATION_KEY
 from executorch.backends.qualcomm.quantizer.quantizer import (
     get_16a8w_qnn_ptq_config,
-    get_default_8bit_qnn_ptq_config,
-    QuantizationConfig,
-)
-from executorch.backends.qualcomm.quantizer.utils import (
+    get_8a8w_qnn_ptq_config,
     get_ptq_per_channel_quant_config,
-    QUANT_ANNOTATION_KEY,
+    QuantizationConfig,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from torch.ao.quantization.quantizer import (
@@ -113,7 +111,7 @@ def annotate_matmul_input1(node: Node, quantization_config: QuantizationConfig):
     # Annotate 16a8w for matmul op to get better performance
     quantization_config_16a8w = get_16a8w_qnn_ptq_config()
     # Annotate 8a8w for second input of matmul until past_kv_cache
-    quantization_config_8a8w = get_default_8bit_qnn_ptq_config(act_symmetric=True)
+    quantization_config_8a8w = get_8a8w_qnn_ptq_config(act_symmetric=True)
     for node in gm.graph.nodes:
         if node.op == "call_function" and node.target == torch.ops.aten.matmul.default:
             if "nn_module_stack" in node.meta:

backends/qualcomm/quantizer/custom_annotation.py

@@ -0,0 +1,108 @@
+import torch
+from torch.ao.quantization.observer import UniformQuantizationObserverBase
+
+
+# TODO move to torch/ao/quantization/observer.py.
+class PerChannelParamObserver(UniformQuantizationObserverBase):
+    """
+    Minimize quantization loss caused by outlier via linear search. More details can be found at https://arxiv.org/pdf/2209.13325
+    """
+
+    def __init__(
+        self,
+        ch_axis=0,
+        use_mse=True,
+        steps=100,
+        dtype=torch.int8,
+        qscheme=torch.per_channel_symmetric,
+        reduce_range=False,
+        quant_min=None,
+        quant_max=None,
+        factory_kwargs=None,
+        eps=torch.finfo(torch.float32).eps,  # noqa: B008
+        is_dynamic=False,
+        **kwargs,
+    ) -> None:
+        super().__init__(
+            dtype=dtype,
+            qscheme=qscheme,
+            reduce_range=reduce_range,
+            quant_min=quant_min,
+            quant_max=quant_max,
+            factory_kwargs=factory_kwargs,
+            eps=eps,
+            is_dynamic=is_dynamic,
+            **kwargs,
+        )
+
+        factory_kwargs = torch.nn.factory_kwargs(factory_kwargs)
+        self.register_buffer("min_val", torch.tensor(float("inf"), **factory_kwargs))
+        self.register_buffer("max_val", torch.tensor(float("-inf"), **factory_kwargs))
+        self.ch_axis = ch_axis
+        self.use_mse = use_mse
+        self.steps = steps
+        self.calibrated = False
+
+    def to_ch_axis(self, x):
+        axis_order = list(range(len(x.size())))
+        axis_order[self.ch_axis], axis_order[0] = 0, self.ch_axis
+        return torch.flatten(x.permute(axis_order), start_dim=1)
+
+    def mse(self, pred, expect):
+        loss = (pred - expect).abs().pow(2)
+        return self.to_ch_axis(loss).mean(1)
+
+    def cosine(self, pred, expect):
+        target = torch.ones(pred.shape[self.ch_axis])
+        pred_n = self.to_ch_axis(pred).reshape(pred.shape[0], -1)
+        expect_n = self.to_ch_axis(expect).reshape(expect.shape[0], -1)
+        return torch.nn.CosineEmbeddingLoss()(pred_n, expect_n, target)
+
+    def loss_fn(self, x, new_min, new_max):
+        scale, offset = self._calculate_qparams(new_min, new_max)
+        x_q = torch.fake_quantize_per_channel_affine(
+            x,
+            scale.data,
+            offset.data.int(),
+            self.ch_axis,
+            self.quant_min,
+            self.quant_max,
+        )
+        return self.mse(x_q, x) if self.use_mse else self.cosine(x_q, x)
+
+    def line_search(self, x):
+        x_min, x_max = torch.aminmax(self.to_ch_axis(x), dim=1)
+        x_range = torch.max(x_min.abs(), x_max)
+        optimal_loss = torch.zeros_like(x_min) + 1e9
+
+        # check which clip range could produce smallest loss
+        for i in range(1, self.steps + 1):
+            thres = x_range / self.steps * i
+            current_loss = self.loss_fn(x, -thres, thres)
+            x_min = torch.where(current_loss < optimal_loss, -thres, x_min)
+            x_max = torch.where(current_loss < optimal_loss, thres, x_max)
+            optimal_loss = torch.min(current_loss, optimal_loss)
+
+        return x_min, x_max
+
+    def forward(self, x_orig):
+        # since params are static, one calibration is enough
+        if not self.calibrated:
+            x = x_orig.detach().to(self.min_val.dtype)
+            self.min_val, self.max_val = self.line_search(x)
+            self.calibrated = True
+
+        # return fake-quant result for saturating outliers
+        scale, zero_point = self._calculate_qparams(self.min_val, self.max_val)
+        return torch.fake_quantize_per_channel_affine(
+            x_orig,
+            scale.data,
+            zero_point.data.int(),
+            self.ch_axis,
+            self.quant_min,
+            self.quant_max,
+        )
+
+    @torch.jit.export
+    def calculate_qparams(self):
+        return self._calculate_qparams(self.min_val, self.max_val)